Interpretable artificial intelligence modeling of pre-emergence herbicide bioactivity in weakly weathered soils for optimized dose recommendations, part I: diclosulam

Abstract

Conventional herbicide recommendations seldom consider soil physicochemical attributes beyond texture, overlooking key factors that govern bioavailability and environmental fate. This study presents an integrated framework for optimizing the doses of pre-emergence herbicides in eight highly heterogeneous, weakly weathered soils derived from Andean sediments. Bioassays with Sorghum bicolor were modeled using extreme gradient enhancement, interpreted using Shapley additive explanations, and refined using symbolic parametric optimization. The approach replaces soil-specific dose–response curves with a single generalizable function calibrated on continuous soil descriptors. A three-dimensional response surface relating herbicide dose and soil enabled intuitive visualization of non-linear interactions and supported the derivation of an optimized and interpretable dose–response function for herbicide recommendation. Soil , electrical conductivity, sulfate, and magnesium emerged as the main predictors of diclosulam bioactivity. Although the dose required for 90% weed control varied between soils, optimized rates remained below active ingredient ha−1, representing reductions greater than 70% relative to commercial label recommendations. The final model, interpretable and parsimonious, predicts weed control as a function of herbicide dose and soil , with flexibility to incorporate additional variables. These results highlight the potential for soil-adaptive herbicide management by identifying site-specific attributes important for optimizing applications, thus improving efficiency, reducing environmental contamination risk, and advancing sustainable and precision agriculture in alignment with the United Nations 2030 Agenda for Sustainable Development.